Important and unimportant patient characteristics for tremor prevalence and tremor irregularity in dystonia
Sinem Balta Beylergil1,2, Laura Scorr3, Adam Cotton3, H. A. Jinnah3,4, and Aasef G. Shaikh2,5
for the Dystonia Coalition Investigators
1Department of Biomedical Engineering, Case Western Reserve University, Cleveland OH, 2Department of Neurology, VA Medical Center, Cleveland OH, 3Department of Neurology, Emory University, Atlanta GA, 30322, 4Department of Human Genetics, Emory University, Atlanta GA, 30322, 5Department of Neurology, Case Western Reserve University, Cleveland OH, and Department of Neurology, University Hospitals, Cleveland OH
Objective: (1) To evaluate the importance of dystonia features that can predict concurrent tremor prevalence and tremor irregularity. (2) To cluster dystonia cases based on important data-driven features using a large, multi-institutional cohort of 2362 patients, and adopting state of the art feature selection and clustering methods of machine learning.
Background: Essential tremor affects 7 million people in the United States and results in substantial social and physical disability. Dystonia is the third most common movement disorder affecting more than 3 million people worldwide. Although essential tremor and dystonia are viewed as distinct disorders, they are closely related. One of the most controversial relationships between dystonia and essential tremor involves the concept of dystonic tremor. Essential and dystonic tremor both are characterized by rhythmic oscillations, but essential tremor has regular waveforms, while dystonic tremor waveforms are irregular. The dystonia features that increase the likelihood of concurrent tremor are unclear. Dystonia traits associated with irregular tremor are also not well defined. Such identification will allow accurate diagnosis and customized therapy for these common movement disorders.
Methods: We used a permutation-based feature selection algorithm to evaluate various dystonia attributes and select the relevant ones to be used in predicting tremor prevalence and irregularity. We performed clustering analyses to group 2362 dystonia patients into clusters with similar characteristics using an agglomerative hierarchical clustering algorithm.
Results: The first feature selection analysis indicated that body part affected by dystonia provides the most useful information for predicting tremor prevalence. Duration of dystonia, total Global Dystonia Rating Scale score, and age at dystonia onset also play a significant role in determining whether dystonia and tremor coexist. With these parameters, a random forest classifier (RFC) was able to classify test data with 69% accuracy. The clustering analysis yielded 4 distinct clusters with 16.33%, 30.6%, 62.31% and 67.34% tremor prevalence rates. Additionally, tremor irregularity was found to be sensitive to the extent to which dystonia and tremor locations overlap. RFC was able to predict irregularity with 79% accuracy, and clustering analysis formed 4 distinct clusters with 28.07%, 75.96%, 79.37%, and 84.15% irregular tremor rates. The second feature selection analysis showed that investigator is also an important feature that discriminates between regular and irregular tremor. Handedness, gender, and race were found unimportant for tremor prevalence and irregularity prediction.
Conclusions: We identified the most relevant dystonia traits for predicting concurrent tremor prevalence and irregularity using modern machine learning methods. Our results also exemplify the use of machine learning methods in understanding the relationship between subtypes of heterogeneous movement disorders.